FactionSim
FactionSim is a managed multi-agent civilization simulator set in a Divergent-inspired dystopia.
Core trick
A World Director advances ticks and activates a subset of 15 LLM agents, each acting from its own memory and faction loyalty.
Description
FactionSim is a managed multi-agent civilization simulator set in a
Divergent-inspired dystopia. Instead of scripting characters, we let Gemini reason
as 15 distinct citizens, leaders, and one hidden "Divergent" agent — each with
their own personality, memories, and faction loyalties — then watch a society
emerge.
README excerpt
# FactionSim FactionSim is a managed multi-agent civilization simulator inspired by the faction-based social systems of Divergent. It is not a player-controlled game. It is an observable AI society where citizens, leaders, and one hidden Divergent agent interact inside a pressured dystopian world. The core idea: > What if faction ideology was executable, every citizen had memory and emotion, and one hidden Divergent could destabilize an entire AI society? ## What we are building FactionSim simulates a society made of ideological factions, spatial zones, world metrics, and autonomous agents. Agents make choices based on faction loyalty, personality, relationships, memory, local events, and social pressure. A World Director advances the simulation in ticks, selectively activates agents, records replay data, and eventually feeds the raw event stream into a narrator that produces episode summaries. The MVP target is intentionally small but deep: - Five major faction zones plus factionless space - A small population of agents with distinct personalities and loyalties - One hidden Divergent citizen who does not fit a single ideology - Emergent rebellion pressure instead of scripted story beats - A live map/dashboard and replay viewer for observing the civilization ## Current phase Current target: Phase 1 — project scaffold. This phase sets up the monorepo, frontend shell, shared schemas, generated seed data, planning docs, and repo conventions. AI model calls and Google A